Ertas for Sentiment Analysis

The Challenge

Understanding customer sentiment is critical for product decisions, brand monitoring, and customer experience management. Yet generic sentiment analysis tools consistently fail on industry-specific language. A phrase like 'this policy is aggressive' is negative in customer support but positive in investment analysis. Sarcasm, domain jargon, and cultural context create layers of ambiguity that off-the-shelf models handle poorly, leading to misleading dashboards and misguided business decisions.

The problem compounds at scale. When organizations process thousands of reviews, support tickets, survey responses, and social media mentions daily, even a small error rate translates into significant misclassification volumes. A 10% error rate on 5,000 daily customer interactions means 500 misread signals per day, enough to distort trend analysis, delay escalations, and mask emerging product issues. Fine-grained sentiment categories beyond simple positive/negative, such as frustrated, confused, delighted, or indifferent, require domain-specific training data that generic models simply do not have.

The Solution

Ertas lets teams fine-tune sentiment models on their own labelled data, so the model learns the sentiment categories and language that matter to their business. In Ertas Studio, analysts upload labelled examples as JSONL (customer reviews tagged with sentiment, support tickets tagged with urgency and emotion, survey responses tagged by theme and experience) and train an open model on them through a visual canvas. Data Craft, Studio's guided dataset builder, lets the team edit labels, keep versions of the dataset and approve examples before training.

One fine-tuned model can return several dimensions at once, such as overall sentiment, emotional tone, urgency and topic, in a single response. The trained model downloads from Hub as GGUF for Ollama or llama.cpp, or as safetensors for vLLM, and runs on your own servers for real-time dashboard feeds or batch jobs over historical data, so the customer text it reads stays in-house. Training runs on Ertas's cloud GPUs and the training examples are uploaded to Ertas, so strip names, email addresses and account numbers from them first. As language shifts with new slang, product names and references, the team adds fresh examples and trains again. Teams currently paying a frontier model API to tag customer text can also ask about the adaptive model endpoint, where a custom model trained on their own requests takes over a growing share of them once it passes their evals.

Key Features

Example Workflow

A SaaS company wants to track customer sentiment across support tickets, NPS surveys and app store reviews. Its product team exports 30,000 labelled examples across five sentiment categories (delighted, pleased, neutral, frustrated, churning), strips names, email addresses and account details, and uploads the JSONL dataset to Ertas Studio. They fine-tune Qwen 2.5 7B to classify text across three dimensions (sentiment category, feature area and urgency) and check it against a held-out set the product team labelled by hand. The model is exported from Hub and runs on the company's own servers, where the data pipeline sends it every incoming customer touchpoint. Sentiment data flows into their Looker dashboard, giving the product team daily visibility into how each feature release lands with customers. When a firmware update triggers a spike in frustrated sentiment in the hardware-reliability category, the team sees it on the next day's dashboard instead of waiting for the monthly NPS report.